SynWeaver Framework Improves Web Agents' Ability to Generalize Across Unfamiliar Sites

Autonomous web agents often struggle when encountering websites they haven't been trained on, a limitation that stems from insufficient site-specific guidance. Researchers have now proposed SynWeaver, a framework that addresses this generalization gap by automatically learning how websites work and generating realistic training tasks and trajectories without extensive human annotation.

The Core Problem#

Current data synthesis methods for web agents rely on exploratory approaches to reduce manual labeling work, but they suffer from critical shortcomings. Most exploration strategies fail to discover the full range of features and functionality a website offers, and without deep knowledge of a site's structure and mechanics, they tend to propose unrealistic or impossible tasks. These "hallucinated" tasks then undermine the quality of subsequent training data, making agents less effective when deployed to new domains.

How SynWeaver Works#

The framework operates through three coordinated stages. First, it performs methodical website exploration to create a comprehensive website map that captures diverse page states and the interactions available on each page. This foundational knowledge serves as website-specific supervision for training models that understand the UI landscape.

Second, SynWeaver extracts page-level and transition-level supervision from the map, enabling a UI-aware model to propose tasks with stronger grounding in actual site capabilities. This reduces the tendency to imagine impossible actions.

Third, the system performs collaborative task-trajectory synthesis, where proposed tasks and their execution paths are updated together when inconsistencies emerge. The framework then verifies and repairs the collected data to ensure both feasibility and semantic alignment.

Demonstrated Performance#

Testing on benchmark environments—WebArena and WebVoyager—shows that SynWeaver consistently surpasses competing synthesis baselines. The resulting supervision improves agent performance not only on sites seen during training but also generalizes more effectively to new, out-of-domain websites.

Why It Matters#

The ability to automatically generate high-quality training data from website structure could significantly accelerate the development of more capable and adaptable web agents, potentially reducing the burden of manual annotation while expanding their practical utility across diverse online platforms.